Recommendation data enhancement method based on fine tuning large language model

Abstract

The invention discloses a recommendation data enhancement method based on a fine-tuning large language model, and the method comprises the following steps: S1, determining an optimization target of a data enhancement task and designing a corresponding instruction template for a core task in combination with user-article interaction characteristics and auxiliary information in a recommendation system scene; s2, adjusting parameters of the large language model according to related contents of the data enhancement task by adopting a lightweight fine tuning technology to generate data highly aligned with the core task so as to finely tune the large language model; s3, generating new user-article interaction data by fine tuning the large language model, and supplementing article feature information and enhanced data of a project summary; and S4, integrating the enhanced data with the original data of the recommendation model to optimize the training of the recommendation model. According to the method, user-project interaction is enhanced, project attributes are enriched, a high-quality project summary is generated, the problem of data sparsity is effectively relieved, and the generalization ability and recommendation effect of a recommendation system are improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of recommendation systems and fine-tuning large language models, and in particular to a recommendation data enhancement method based on fine-tuning a large language model. Background Art

[0002] Recommendation systems have become a key component of modern digital platforms, playing an indispensable role in e-commerce, social media, online streaming and other fields. With the rapid growth of data volume and the increasing complexity of user needs, recommendation systems are faced with problems such as data sparsity and cold start, which seriously hinder the improvement of recommendation performance. Traditional data augmentation methods include graph-based, negative sampling and diffusion models. Although these methods improve recommendation performance, they are limited by complexity and model expression ability, especially when dealing with complex contexts and multimodal data. Compared with traditional methods, large language models (LLMs) can generate rich user-item interaction data and attribute descriptions from a small amount of information based on their powerful language understanding and generation capabilities. Such methods usually guide large models to generate diverse interaction records and attribute information through prompt engineering and instruction design, enriching the training data of recommendation systems. Although the problem of data sparsity is fundamentally alleviated, large language models are often accompanied by problems such as output uncertainty, instruction design complexity, and knowledge gaps in the recommendation field while generating rich semantic content. Summary of the invention

[0003] In view of the above shortcomings, the present invention proposes a recommendation data enhancement method based on fine-tuning a large language model, which effectively alleviates the data sparsity problem and improves the generalization ability and recommendation effect of the recommendation system by enhancing user-item interaction, enriching item attributes, and generating high-quality item summaries.

[0004] To achieve the above object, the present invention provides the following technical solution: a recommendation data enhancement method based on fine-tuning a large language model, comprising the following steps:

[0005] S1: Focusing on the core tasks in the recommendation system of the recommendation model, combined with the user-item interaction characteristics and auxiliary information in the recommendation system scenario, determine the optimization goals of the data enhancement tasks and design the instruction templates for the corresponding data enhancement tasks;

[0006] S2: Use lightweight fine-tuning techniques to adjust the parameters of the large language model for content related to the data augmentation task to generate high-quality data that is highly aligned with the core task, thereby fine-tuning the large language model;

[0007] S3: Generate new user-item interaction data, supplement item feature information, and enhance project summary data by fine-tuning the large language model;

[0008] S4: Integrate the enhanced data with the original data of the recommendation model to optimize the training of the recommendation model.

[0009] As an improvement, the optimization objectives of the data augmentation task in step S1 include user-item interaction enhancement, item attribute completion, and item summary generation.

[0010] As an improvement, the instruction template in step S1 includes an instruction input module and an instruction output module. The instruction input module includes a task instruction module and a task input module, and the instruction output module includes a task output module. The task instruction module and the task input module describe the data augmentation task instructions in natural language to guide the large language model to generate data and output it through the task output module.

[0011] As an improvement, in step S2, the low-rank adaptation fine-tuning technique is adopted to only fine-tune the user-item interaction generation strategy, item attribute expansion ability, and item summary generation method to reduce the computational overhead and ensure the quality and consistency of the generated data.

[0012] As an improvement, the low-rank adaptation fine-tuning technique adopts multi-objective fine-tuning, adds a trainable parameter matrix to each layer of the Transformer structure in the pre-trained model of the recommendation model, and realizes task alignment by jointly optimizing the supervised objective and the generation objective, including the following objectives:

[0013] The fine-tuning of the user-item interaction generation strategy fills in the sparse interaction information in the original data by generating new user-item interaction edges. The optimization objective is expressed by the following formula:

[0014]

[0015] where x represents the input, y represents the output, Z represents the training set, y t represents the t-th token in the output, y<t represents all tokens before the t-th token, Φ is the original parameter of M, and Θ1 is the rank decomposition matrix parameter representing the new interaction generation;

[0016] The fine-tuning of the item attribute expansion ability makes the item attributes contain more descriptive information by expanding the features of the item. The optimization objective is expressed by the following formula:

[0017]

[0018] where a t represents the attribute feature of the t-th token, and Θ2 is the rank decomposition matrix parameter for generating or expanding item attributes;

[0019] The fine-tuning of the item summary generation method generates a more representative summary for each item to facilitate more accurate recommendations. The optimization objective is expressed by the following formula:

[0020]

[0021] where p t represents the summary of the tth token, Θ3 is the rank decomposition matrix parameter used to generate the item summary;

[0022] The final learning objective is calculated as:

[0023]

[0024] Among them, λ1, λ2, and λ3 are hyperparameters that balance the importance of each objective.

[0025] As an improvement, in step S3, the user-item interaction data generated by fine-tuning the large language model and inferring potential interests from the user's historical behavior is expanded to reflect the user's real interest distribution and behavior pattern, which is formalized as:

[0026]

[0027] in It is LLM according to the input prompt The obtained positive and negative interaction samples.

[0028] As an improvement, the item feature information supplemented in step S3 is generated based on the existing auxiliary information of the item, and additional descriptive features are generated. The generated features include a more detailed description. The item attribute enhancement is expressed as:

[0029]

[0030] Among them A i Represents the text attribute of the generated item i, text is the attribute refinement of item i.

[0031] As an improvement, the item summary in step S3 generates a high-quality natural language description for the item, outlining the content characteristics and main attributes of the item. The specific process is summarized as follows:

[0032]

[0033] Where P i Represents the textual summary of the generated item i.

[0034] As an improvement, in step S4, the enhanced data and the original data are integrated in a unified format to form a new training set, and adapted to the needs of different recommendation models so as to be seamlessly integrated into a variety of existing recommendation models to optimize the training of the recommendation models.

[0035] Compared with the prior art, the advantages of the present invention are:

[0036] (1) The present invention makes full use of the excellent capabilities of the large language model in semantic understanding and high-quality data generation, and introduces it into the field of data enhancement for the recommendation system. With the powerful text generation and context analysis capabilities of the large language model, it can generate richer, more diverse and highly reliable user-item interaction data and auxiliary information, significantly improving the coverage and accuracy of the recommendation data, thereby optimizing the performance of the recommendation system in data sparsity and cold start scenarios;

[0037] (2) The present invention designs sophisticated fine-tuning targets to align the large language model with the specific data augmentation tasks of the recommendation system. The fine-tuning process not only considers the unique needs of the recommendation system scenario, but also accurately optimizes tasks such as user behavior, preference inference, and item attribute expansion. The generated data better meets the needs of the actual application scenarios of the recommendation system and can better reflect user interests and item characteristics, thereby significantly improving the relevance and effectiveness of the recommendation results and enhancing user experience and recommendation satisfaction.

[0038] (3) The present invention designs customized fine-tuning instructions, which guide the large language model to generate target data through clear generation goals and multi-level task guidance. Compared with traditional manual rules and fixed template methods, this mechanism has higher flexibility and automation capabilities. By reducing manual intervention and development complexity, it not only greatly reduces labor costs, but also improves the generation efficiency of data enhancement, enabling the recommendation system to quickly adapt to changing application scenarios and dynamic data requirements;

[0039] (4) The method proposed in this invention is not only applicable to a specific model, but also model-agnostic, that is, it can be seamlessly integrated into various existing recommendation models to enhance the performance of these models;

[0040] (5) In the present invention, by clarifying the optimization goal, the data generated by each task is ensured to be more targeted and practical, and a refined template is designed to make the generation process more controllable and stable;

[0041] (6) During the training process, the present invention uses quantized low-rank adaptation to perform efficient parameter fine-tuning, thereby greatly reducing the number of trainable parameters and accelerating the training process;

[0042] (7) The present invention uses low-rank adaptive lightweight fine-tuning technology to adjust only a small number of model parameters to reduce computational overhead while ensuring the quality and consistency of data generated by the model. The core idea of ​​low-rank adaptive lightweight fine-tuning technology is to decompose the changes in model parameters into a low-rank matrix form and only update the decomposition matrix to complete the fine-tuning of specific tasks. This method avoids direct adjustment of all model parameters, effectively reduces the amount of computation and storage requirements, and maintains the accuracy and consistency of data generation while reducing computational costs, meeting the high-quality data requirements of the recommendation system. This method can be easily extended to different task requirements without large-scale changes to the model structure, thereby improving overall applicability.

[0043] (8) By jointly optimizing multiple task objectives, the model's comprehensive adaptability to the scenario requirements of the recommendation system is improved. Through parameter sharing and multi-objective optimization, parameter redundancy during fine-tuning is reduced. At the same time, performance and efficiency are taken into account. The objective functions designed for different tasks make the generated data more in line with actual recommendation needs.

[0044] (9) By inferring potential interests and generating new interaction data, the scale of user-item interaction data is significantly expanded. The generated data is based on a fine-tuned large language model, which has higher semantic understanding ability and context consistency, improving the quality of recommendations.

[0045] (10) By generating additional features, the feature dimension of items is significantly expanded, providing more comprehensive input information for the recommendation model. The generated feature description has strong semantic relevance and adaptability and can be widely used in different recommendation scenarios.

[0046] (11) By generating high-quality natural language summaries, the semantic expression ability of item information can be significantly improved, enabling the recommendation system to better understand the characteristics of items. The summary can clearly and concisely summarize the core content of the item, making it easier for users to quickly understand the key information of the item.

[0047] (12) The newly generated supplementary features and project profiles are added to the original data according to the established fields without changing the basic organizational structure of the original data. For different recommendation scenarios, the ratio of original data to enhanced data is dynamically adjusted so that it can better balance diversity and relevance during training. In the process of integrating data, the basic characteristics of the original data are retained to ensure that the enhanced data has a high degree of consistency with the original data after being added, so that the recommendation model can seamlessly adapt to the input after data enhancement. Through the model-agnostic design, there is no need to adjust the data format or characteristics separately for different recommendation algorithms, reducing the development workload and the cost of model migration. By adopting a dynamic adjustment strategy, the ratio of enhanced data can be flexibly configured according to actual needs, so that the recommendation system can better adapt to different data distributions and scenario requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0049] Figure 1 A brief flowchart of a recommendation data augmentation method based on fine-tuning a large language model;

[0050] Figure 2 A detailed flowchart of a recommendation data augmentation method based on fine-tuning a large language model;

[0051] Figure 3 Schematic diagram for lightweight fine-tuning of large language models;

[0052] Figure 4 Enhanced graphs for model-agnostic recommendations. DETAILED DESCRIPTION

[0053] like Figures 1 to 4 As shown, a recommendation data enhancement method based on fine-tuning a large language model includes the following steps:

[0054] S1: Determine the fine-tuning target:

[0055] Combined with the actual application scenarios of the recommendation system, we analyze the interaction characteristics between users and items, the user's historical behavior (clicks, purchases, ratings, etc.), the user's interest change patterns (long-term interests and short-term interests), and the availability of auxiliary information (item descriptions, classification labels, etc.). Based on the core tasks of the recommendation system, we determine the following optimization goals:

[0056] (1) User-item interaction enhancement: Generate new interaction records consistent with the user's historical behavior to enrich the user's behavior sequence;

[0057] (2) Item attribute completion: Based on the existing item auxiliary information, more fine-grained descriptive features are added to enhance the expressiveness of item features;

[0058] (3) Project summary generation: Generate a high-quality natural language summary for the project, summarizing its main content and key attributes;

[0059] Build instruction templates for various data augmentation tasks, and describe fine-tuning goals in natural language to clearly guide the model to generate data;

[0060] Project outline generation instruction template:

[0061]

[0062] User-item interaction enhancement instruction template:

[0063]

[0064]

[0065] Item Attribute Completion Instruction Template:

[0066]

[0067] S2: Fine-tune the large language model:

[0068] The capabilities of the large language model are customized to meet the requirements of specific tasks. A lightweight adjustment strategy is adopted to achieve efficient multi-objective fine-tuning. The central premise of this strategy is that contemporary language models usually have a large number of parameters, but the information of these parameters may be concentrated in lower intrinsic dimensions. Therefore, by adjusting only a small part of the parameters, a performance similar to that of fully fine-tuning the entire model can be achieved. A lightweight fine-tuning technique called LoRA is adopted on lamafactory. The core idea of the LoRA method is to freeze most of the parameters of the pre-trained model and introduce trainable rank decomposition matrices only in each layer of the Transformer architecture.

[0069] Specifically, task alignment is achieved through the following optimization of the following fine-tuning objectives:

[0070] User-Item Interaction Generation Strategy Fine-tuning fills in sparse interaction information in the original data by generating new user-item interaction edges. The optimization objective is represented by the following formula:

[0071]

[0072] where x represents the input, y represents the output, and Z represents the training set. y t represents the t-th token in the output, and y<t represents all the tokens before the t-th token. Φ is the original parameter of M, and Θ1 is the parameter of the rank decomposition matrix used to generate new interactions;

[0073] Item Attribute Expansion Ability Fine-tuning: Expand the features of the item to include more descriptive information. The optimization objective is represented by the following formula:

[0074]

[0075] where, a t represents the attribute feature of the t-th token, and Θ2 is the parameter of the rank decomposition matrix used to generate or expand item attributes;

[0076] Summary Generation Method Fine-tuning: Generate a more representative summary for each item for more accurate recommendation. The optimization objective is represented by the following formula:

[0077]

[0078] where, pt represents the summary of the tth token, Θ3 is the rank decomposition matrix parameter used to generate the item summary;

[0079] The final learning objective is calculated as:

[0080]

[0081] Among them, λ1, λ2, and λ3 are hyperparameters that balance the importance of each objective;

[0082] S3: Generate augmented data:

[0083] In the face of the sparsity of interaction data, the fine-tuned LLM is used as a sampler to generate paired interaction training data from the perspective of natural language. In this way, potential supervisory signals can be mined from the original data, and contextual knowledge (such as year, genre, etc.) can be integrated into user-item interactions to help better understand user preferences and behaviors. This method is particularly suitable for users and items that lack sufficient historical interaction records, because LLM has been aligned with the recommendation task through fine-tuning and can generate data that is more consistent with the actual interaction distribution. Specifically, first select the items of each user's historical interaction and combine the user's auxiliary information (such as browsing history, rating habits, etc.). Then, introduce an item candidate pool Cu = {i3,i7,…,i n}, which contains items that have interacted with all users less than 10 times in the original dataset. Since LLM cannot process all items directly, this candidate pool is selected to limit the input scale. Then, the user's historical interactions, auxiliary information, and item candidate pool are input into LLM together. During the reasoning process of LLM, it randomly outputs pairs of items that users may like. Or don't like These generated interaction data not only enrich the original sparse interaction matrix, but also better reflect the user's real interest distribution and behavior pattern. The process of enhancing user-item interaction is formalized as follows:

[0084]

[0085] in It is LLM according to the input prompt Positive and negative samples obtained from the candidate pool Cu;

[0086] Faced with the lack of textual information on the item attributes of the original dataset, a method based on fine-tuning LLM is used to enrich the item attribute information by taking advantage of its powerful generation ability. LLM has a huge knowledge base and can understand and generate diverse information related to items. Therefore, LLM can be guided to expand and supplement the missing or incomplete item attributes in the original dataset. Specifically, each item is first provided with auxiliary information (such as title, description, category, etc.) already in the dataset as prompts. These prompt information is input into the fine-tuned LLM, and the model generates additional item attributes based on these prompts. These generated attributes may include more detailed descriptions, additional category labels, related topics, overviews of user reviews, etc., which do not exist or are not complete in the original dataset. Formally, the item attribute enhancement based on LLM is expressed as:

[0087]

[0088] A i Represents the text attribute of the generated item i, text prompt is the attribute refinement of item i;

[0089] The project summary is a comprehensive characterization of the overall characteristics and attributes of the project. It not only describes the main content of the project, but also clarifies the specific types of users that the project is suitable for attracting, and shows the project characteristics and qualities that are consistent with the preferences and interests of these users. An accurate and detailed project summary can help the recommendation system better match users and projects, thereby providing more personalized and accurate recommendations. When constructing a project summary, how to generate high-quality summary information that is highly matched with user needs has always been a challenge. To this end, the method of fine-tuning LLM is adopted to use its powerful ability in natural language generation to create more descriptive and attractive project summaries. Based on these research results, a simplified input prompt is designed and used as part of the LLM input to clearly define the function of LLM in generating project summaries. This prompt provides clear guidance for LLM to ensure that the generated project summary can fully and accurately reflect the main content and unique attributes of the project. The specific process is outlined as follows:

[0090]

[0091] P i The text summary of project i is generated. Specifically, the input prompt only requires basic information about the project and the task requirements for generating the summary. The fine-tuned LLM is able to generate more complete and detailed project summaries based on these prompts.

[0092] S4: Model-agnostic enhancements:

[0093] The present invention is dedicated to improving the recommendation performance of the model by enhancing the interaction data and item text features in the original dataset. The proposed method is not only applicable to specific models, but also model-agnostic, that is, it can be seamlessly integrated into various existing recommendation models to enhance the performance of these models. Specifically, the method generates enhanced data through fine-tuned LLM, which includes richer user-item interaction records and extended item text features. These enhanced data can provide additional training signals for any recommendation model using the dataset, helping the model to better capture user preferences and item features, thereby making more accurate recommendations. The method is highly flexible for different types of recommendation models. For example, for recommendation models based on collaborative filtering, the model mainly relies on the interaction data between users and items, rather than directly utilizing the text information of the items. Therefore, in this case, the enhancement of item text features may not be necessary, and only the enhanced interaction data can be used. For recommendation models that need to utilize text features (such as content-based recommendation models or hybrid models), the enhanced item text features can provide richer input, which helps the model understand the matching degree between item content and user interests. In practical applications, this can be achieved through the following steps: First, using the enhancement process described above, a dataset including enhanced interaction data and text features is generated. Then, these enhanced data are combined with the original data and input into the existing SOTA recommendation model for training.

[0094] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0095] The units described in some embodiments of the present disclosure may be implemented in software or hardware. The units described may also be arranged in a processor, and the functions described above may be at least partially performed by one or more hardware logic components.

[0096] The above description is only for the best embodiment of the present invention, but it should not be understood as limiting the claims. The present invention is not limited to the above embodiments, and its specific structure is allowed to be changed. All changes made within the scope of protection of the independent claims of the present invention are within the scope of protection of the present invention.

Claims

1. A recommendation data enhancement method based on fine-tuning a large language model, characterized in that: The following steps are involved: S1: For the core tasks in the recommendation system of the recommendation model, combined with the user-item interaction characteristics and auxiliary information in the recommendation system scenario, determine the optimization target of the data enhancement task and design the instruction template of the corresponding data enhancement task; S2: Using lightweight fine-tuning technology to adjust the parameters of the large language model for the content related to the data augmentation task to generate high-quality data that is highly aligned with the core task, thereby fine-tuning the large language model; S3: Generate new user-item interaction data, supplementary item feature information, and enhanced data of project summaries through the fine-tuning of the large language model; S4: Integrate the enhanced data with the original data of the recommendation model to optimize the training of the recommendation model.

2. The method for enhancing recommendation data based on fine-tuning a large language model according to claim 1, characterized in that: The optimization objectives of the data enhancement task in step S1 include user-item interaction enhancement, item attribute completion, and project summary generation.

3. The method for enhancing recommendation data based on fine-tuning a large language model according to claim 2, characterized in that: The instruction template in step S1 includes an instruction input module and an instruction output module. The instruction input module includes a task instruction module and a task input module. The instruction output module includes a task output module. The task instruction module and the task input module enhance the task instructions through natural language description data to guide the large language model to generate data and output it through the task output module.

4. The method for enhancing recommendation data based on fine-tuning a large language model according to claim 1, characterized in that: In step S2, low-rank adaptive fine-tuning technology is used to fine-tune only the user-item interaction generation strategy, item attribute extension capability, and project summary generation method to reduce computational overhead and ensure the quality and consistency of generated data.

5. The method for enhancing recommendation data based on fine-tuning a large language model according to claim 4, characterized in that: The low-rank adaptation fine-tuning technology adopts multi-objective fine-tuning, adds a trainable parameter matrix to each layer of the Transformer structure in the pre-trained model of the recommendation model, and achieves task alignment by jointly optimizing the supervision objective and the generation objective, including the following objectives: The user-item interaction generation strategy is fine-tuned by generating new user-item interaction edges to fill the sparse interaction information in the original data. The optimization objective is expressed as follows: where x represents the input, y represents the output, Z represents the training set, y t represents the t-th token in the output, y<t represents all tokens before the t-th token, Φ is the original parameter of M, and Θ1 is the parameter of the rank decomposition matrix used to generate new interactions; Fine-tuning the item attribute expansion capability expands the item's features so that the item attributes contain more descriptive information. The optimization goal is expressed by the following formula: where a t represents the attribute features of the tth token, Θ2 is the rank decomposition matrix parameter used to generate or expand item attributes; The item summary generation method is fine-tuned to generate a more representative summary for each item to facilitate more accurate recommendations. The optimization objective is expressed as follows: where p t represents the summary of the tth token, Θ3 is the rank decomposition matrix parameter used to generate the item summary; The final learning objective is calculated as: Among them, λ1, λ2, and λ3 are hyperparameters that balance the importance of each objective.

6. The method for enhancing recommendation data based on fine-tuning a large language model according to claim 1, characterized in that: In step S3, the user-item interaction data generated by fine-tuning the large language model and inferring potential interests from the user's historical behavior is expanded to reflect the user's real interest distribution and behavior pattern, which is formalized as: in It is LLM according to the input prompt The obtained positive and negative interaction samples.

7. The method for enhancing recommendation data based on fine-tuning a large language model according to claim 1, characterized in that: The supplementary item feature information in step S3 is based on the existing auxiliary information of the item to generate additional descriptive features. The generated features include a more detailed description. The item attribute enhancement is expressed as: Among them A i Represents the text attribute of the generated item i, text is the attribute refinement of item i.

8. The method for enhancing recommendation data based on fine-tuning a large language model according to claim 1, characterized in that: The project summary in step S3 generates a high-quality natural language description of the project, outlining the content characteristics and main attributes of the item. The specific process is summarized as follows: Where P i Represents the text summary of the generated item i.

9. The method for enhancing recommendation data based on fine-tuning a large language model according to claim 1, characterized in that: In step S4, the enhanced data and the original data are integrated in a unified format to form a new training set, and adapted to the requirements of different recommendation models so as to be seamlessly integrated into a variety of existing recommendation models to optimize the training of the recommendation models.

Images